Measuring the Unmeasurable: Ranking One’s Favorite Music, Part 1

I recently updated a data-based discussion of my cinematic “guilty pleasures,” adding a comparison of “most-acclaimed” and “my favorite” films from a given year or years. In so doing, though, I side-stepped the question of determining with something approaching academic rigor just what my favorite films are, relying solely on my gut to select a favorite film or films from each period.

Readers of this website know that I am fascinated by the art and science of measurement, be it the “noirness” of film festivals, Charlie Chan films, the Marvel Cinematic Universe or baseball player performance. Each of these prior analyses, however, is purely objective: all of the data I used are publicly-available, so the only “subjective” decisions I needed to make were selecting which data to compile and what statistical methods to use. And even when I was analyzing data for which I am the lone source – like this gorgeous distribution of iTunes tracks[1] by year and genre – the only decisions I made related to visualization, not personal preference.

Now, at last, I tackle the deceptively simple question of what music, movies, etc. I like most. More to the point, I address how I can most effectively and efficiently derive a “score” for each track or film, so that I can not only rank order them, but aggregate them into, say, albums, artists and genres, overall or by time period.

In the first installment of this seriees, we journey from my first-ever “mixtape” to my initial attempt to assign scores to my favorite music.


While researching my book Interrogating Memory: Film Noir Spurs a Deep Dive Into My Family History…and My Own, I reread a hand-written “Journal” – really just a paper-clipped set of mid-sized lined pieces of paper ripped from a notebook – I began on May 29, 1981, as I was about to finish 9th grade at Harriton High School.

Buried within the June 5, 1981 entry is this:

Went to Ludington [Library in Bryn Mawr, PA]. … Then from there to Sam Goody’s for The Moody Blues Long Distance Voyager and Phil Collins Face Value. Good stuff. … Then it was Mad’s for Kraftwerk Autobahn.

At the time the two record stores sat a short walk from each other on E. Lancaster Avenue in the Philadelphia suburb of Ardmore. Four decades later, I still own these very albums.

A few months earlier, I had convinced my father – not exactly flush with cash much of the time – to buy three albums for me: The CarsThe Cars and Panorama, and, if memory serves, Supertramp’s Breakfast in America. I still have Panorama on vinyl, though I long since replaced the other two with CDs.

In the 12 months before that, meanwhile, I acquired Fleetwood Mac’s Tusk, Peter Gabriel (III) and Steve Winwood’s Arc of a Diver. I still have Tusk and Arc on vinyl, with PGIII replaced by a CD. Finally, in late July or early August 1981, I acquired Foreigner 4…which I foolishly sold a few years later to buy a new issue of Billboard.

I played these 10 albums – and others I owned – on a turntable, complete with mid-sized brown wooden-cabinet speakers, my mother had bought just over four years earlier. Two moves later, it had migrated to my bedroom, where it sat on a low white wooden shelf with my record collection.

Exactly when I received my first Sony Walkman – and when I got the idea to record a set of tracks I like onto a Maxell cassette – I could not tell you. Nor could I tell you exactly what day in August 1981 – it was likely a Saturday, as I was a day camp junior counselor on weekdays – I placed one of the brown speakers on the floor next to my portable cassette recorder, cued up “Spanish Dancer” on Arc of a Diver, and hit Record. I sat in absolute silence as that track – followed by 13 others – recorded monaurally with zero Dolby noise reduction.

At least, I think it was these 14 tracks, in this order:

Side 1

Spanish Dancer                     Steve Winwood

Night Train                           Steve Winwood

Urgent                                    Foreigner

Juke Box Hero                       Foreigner

Moving in Stereo                  The Cars

All Mixed Up                        The Cars

Side 2

Touch and Go                       The Cars

Running to You                    The Cars

In the Air Tonight                Phil Collins

The Voice                              The Moody Blues

Gemini Dream                      The Moody Blues

Sisters of the Moon              Fleetwood Mac

Games Without Frontiers   Peter Gabriel

I Don’t Remember               Peter Gabriel

I have long since lost the cassette and lined insert card on which I wrote track titles and artists. A thorough search of the cardboard box on the floor to my left – containing dozens of mix cassettes and CDs – did not reveal it. It is purely memory that conjures up this list, though it is a very reasonable list.

My memory also says it was a 60-minute cassette, except this version of Side 1 is 32 minutes long, while Side 2 is 36 minutes long. So, unless I cut off two minutes of a track on Side 1 and have too many tracks on Side 2, this was more likely a 90-minute cassette, and I either ran out of tracks to record (VERY unlikely), or I am forgetting tracks.

Either way, as the cardboard box implies, what I creatively called My Stuff was only my first adventure in mix-making. Four months later I filled a 90-minute cassette with 22 tracks recorded from the radio; I called this cassette Stuff Vol. I. Over the next two years, I completed Stuff Vol. II through XI. With the exception of Side 1 of Stuff Vol. IV, a collection of late 60s/early 70s rock I recorded in June 1982,[2] these were “acquisition” mixes – mostly from the radio but sometimes from borrowed records. Indeed, only six of the 188 (or more, see below) unique tracks I recorded onto those mixes were recorded twice: “I Don’t Remember,” “Tainted Love” by Soft Cell, and the four borrowed-record tracks: two by A Flock of Seagulls, “Kids in America” by Kim Wilde and “Escalator of Life” by Robert Hazard and the Heroes.

It is pretty clear where my musical tastes lay then – and now.


I first wrote here about the 300+ mix cassettes, CDs and videocassettes I recorded between August 1981 and August 2016. These mixes contain at least 3,383 unique tracks.

  • Besides My Stuff, I no longer have Stuff Vol. V (from which I recall seven tracks) and one of the two mixes between Stuff Vol. VIII and Stuff Vol. XI. The other one is in a different plastic case with a sticker reading “I92–Stranglers/Fixx/Devo.” As I have no memory of the “missing” cassette – and I played the heck out of these mixes – it is very likely I simply misnumbered them.
  • When I first began to enter mix contents into a spreadsheet in December 1992, I equated studio and live versions of the same track (except when I didn’t), while including tracks from which I only recorded snippets, or which got cut off at the end.

I had actually been rank-ordering my favorite music for years – summoning from my gut then playing personal “top 25 songs” lists since 1980, if not earlier. In January 1990, meanwhile, I commemorated the end of the 1980s by determining – purely through thought and memory – my 40 favorite tracks of that decade; I mistakenly included “Prime Time” by The Tubes, even though Remote Control, its parent album, was released in 1979. “Promised You a Miracle” by Simple Minds was #1, beating out “All Roads Lead to Rome” by The Stranglers. In March, I purchased my first CD player, allowing me to record from CD to cassette, followed in April by my first PC. The former purchase triggered a wave of CD buying, mostly through buy 1, get 11 for one penny deals; I stocked up on “Best of” CDs. I also began prowling through used record, tape and CD stores. With a PC, meanwhile, I now had a spreadsheet program, though not Microsoft Excel.

By December 1992, I had created 88mixes – excluding the two Top 40 of the 1980s cassettes (10 tracks over four sides), a mix I created for my then-girlfriend in 1990 and a mix I created in the summer of 1992 for a woman in whom I was romantically interested. In the days before data compression and MP3s, constructing a mix required you to play every included track in its entirety. Plus, cassettes had only so much room and unless you bought them in bulk, blank ones were precious. Thus, you really had to like a track if you chose to record it onto a mix.

The brainstorm that I had in December 1992, then, was this: the universe of tracks included on my mixes roughly corresponded to my favorite tracks. And by tallying up what tracks appeared most often on a mix, I could both rank my favorite tracks and assign each one a numeric score. Then, by aggregating those scores, I could rank my favorite albums, artists and genres, both overall and by year.

Ranking hypothesis #1: I like every track I recorded onto a mix more than any track I have never recorded onto a mix.

Ranking hypothesis #2: The more mixes on which I have recorded a track, the more I like it.

This proved not to be as straightforward as I had hoped.


In March 1984, I drove some classmates and myself to Washington DC for a Model UN gathering. Anticipating the long drive, I made my first proper mix tapes since My Stuff: Georgetown Survival Mix Vol. I and II. We did not actually stay in Georgetown – where a teenaged Nell (now my wife) then lived – but in a Marriott on Connecticut Avenue NW in Woodley Park. These 43 tracks summarized favored recent album purchases: ABC’s The Lexicon of Love and Spandau Ballet’s True, plus an assortment of records from Roxy Music, Alan Parsons Project, Talking Heads, Genesis, Peter Gabriel, A Flock of Seagulls, U2, Duran Duran, Squeeze, Icehouse, Re-Flex and Real Life. I still did not have a proper cassette recorder – nor could I play cassettes in the “Berger Bus,” my black 1979 Ford Fairmont. So, my tape recorder came along for the ride.

Similar two-cassette “constructed” mixes followed in July 1984, November 1984 and January to March 1985 – with another “acquisition mix” in June 1984. That summer, as I settled into my new room in my mother’s suburban Philadelphia apartment – she had rented my bedroom to a young woman – I was gifted an all-in-one turntable/cassette player/radio/cassette recorder. I spent much of that summer twirling radio dials, seeking tracks to record, creating Summer 1985, Vols. I to VIII.

At the start of my sophomore year at Yale, after 28 mixes and 487+ unique tracks, I constructed a mix – mostly from my record collection, but with some borrowed albums as well – called Stuff and Such Vol. I. I have no idea why I chose this playful title, but in so doing I created the mix-naming convention I would use for the next 16 years.

Well, it was the naming convention I used exclusively, excepting Pseudo Dance Music Vols. I and II in October 1989, beginning with my move to the Boston suburbs at the end of August 1989. In the previous four years, I intertwined first 12 Stuff and Such volumes with Stuff of 1985, Summer 86 Vols. I-III, Summer 1987 Vols. I-III, Stuff of 1987-88, Video Stuff Vols. I-IV, Summer 1988 Vol. I and Washington Vol. I-III. Broadly speaking, I constructed Stuff and Such mixes from my record collection – filling out sides with a few radio/borrowed tracks – and the other mixes from the radio and elsewhere – filling out sides with tracks from my record collection.

Basically, mixes served two purposes at this time:

  1. Acquisition: Though borrowing and recording from the radio and cable music channels new tracks became part of my collection.
  2. Portability: I could play tracks from newly-acquired albums on my Sony Walkman anywhere I wanted.

By the end of August 1989, I had created 52 cassette and four video mixes – averaging seven per year – comprising 919+ unique tracks spread over 1,118+ “slots.” A “slot” is anything recorded onto a mix, so if I record 25 tracks onto a cassette, that is 25 “slots.” During this eight-year period, 157 tracks (17.1%) were recorded twice, while 21 (2.3%) were recorded three times. A track appears multiple times either because I liked it so much, I wanted it to include it on subsequent mixes, or because I first recorded one version (perhaps a 12” remix) from the radio then found a different version – or the video on either MTV or VH1.

Updated Ranking Hypothesis #2: The number of mixes onto which I record a track is positively associated with how much I like that track.

Meanwhile, as I prepared to move to Somerville to enroll in the Harvard Graduate School of Arts and Sciences doctoral program in government, I knew I had a long solo drive to Boston ahead of me. To accompany me on this drive, and presuming I could play these cassettes on the sound system of the U-Haul I drove, I lovingly prepared the 136-track Boston Drive Vols. I-VI. Only nine tracks were being recorded for the first time, including the first two tracks on Side 1 of Boston Drive Vol. I: Overture and Heaven on Their Minds, which open the soundtrack to the 1973 film version of Jesus Christ Superstar; I played that soundtrack incessantly over the next few years. This is the first instance of what I later called the “anchor” track (in this case, tracks) – the very first track recorded on a mix or set of mixes, the one(s) I am the most excited to hear. This would later apply to a) the first track recorded on any mix within a set and b) the final track recorded on that set.

Ranking Hypothesis #3: The tracks I like the most on a mix or set of mixes are the first tracks I record on a cassette/CD – especially the anchor track – and the final track I record.

Ranking Hypothesis #3a: RH3 is not true before August 1989.

Ranking Hypothesis #3b: RH3 is sometimes true between September 1989 and February 1992.

Ranking Hypothesis #3c: RH3 is always true after February 1992.

In essence, the Boston Drive mixes were a compendium of those tracks I always fast-forwarded or rewound to hear over the previous eight years. And for the first time, I began to think about how I ordered tracks. Up until now I had either been at the whim of disc/video jockeys or had recorded artist “blocks” – a group of ABC tracks followed by a group of Spandau Ballet tracks followed by…you get the idea. There was no particular ordering of tracks within each artist – or really at all.

This began to change with Boston Drive, even my thought process was no more complicated than “put ‘rocking’ tracks on Side 1 and ‘mellow’ tracks on Side 2,” with “rocking” and “mellow” loosely defined. This was how I constructed Vols. I-III. Side 1 of Boston Drive Vol. IV contains 10 Genesis tracks, while Side 2 opens with “The Chamber of 32 Doors” followed by seven instrumental tracks. Boston Drive Vols. V and VI are essentially the leftovers, with the line between “rocking” and “mellow” nearly obliterated. The final track – still one of my 10 or 15 favorites – is “Darkness,” the last track on The Police’s Ghost In The Machine.

The Boston Drive mixes were such a revelation I returned to them repeatedly over the next few years. They also allowed me essentially to ignore the previous 52 cassette mixes, though I still watched the four video cassettes sometimes; what I do not remember is whether I had physically left them behind in my mother’s suburban Philadelphia apartment.

Meanwhile, shortly after moving to Somerville, I borrowed an apartment-mate’s CD of Camper Van Beethoven’s Our Beloved Revolutionary Sweetheart. I had fallen in love with “One of These Days” after hearing it played on WFNX in Washington, though I never got the chance to record it. That track – still one of my 100 favorites – not only opened Stuff and Such Vol. XIII, it signaled the end of the first wave of mix making. I have not recorded a single track from the radio since then, for example, though I did once fill an entire videocassette from VH1 Classic, retroactively designating a set of videos as Video Stuff 2002-03. Also, I now focused solely on constructing cassette mixes based on a mix of newly-acquired tracks – whether from CDs, vinyl albums or cassettes, or culled from other collections – and “repeat” tracks, the ones I wanted to hear again after recording them on an earlier mix. I still grouped tracks by artist – devoting two cassettes to Genesis and one to Roxy Music – and gave little thought to how one track flowed into the next, but the first faint glimmers of the strict mix-construction rules I later followed are there.

This demarcation of mixes into “before Boston Drive” and “Boston Drive and later,” however, meant that when I began to enter track name, artist and mix name into my PC spreadsheet I began with the Boston Drive mixes. I was daunted enough by the thought of entering data from 32 mixes – 36 counting the tangential 80s and romantic mixes; entering data from the previous 56 mixes went above and beyond.

Moreover, I was not satisfied with a simple tally of how many of the 749 slots each of the 512 tracks filled; two-thirds (341) occupied only one slot, while 39 occupied three slots, six occupied four slots and five – “Zamba” by Bryan Ferry, “Save Me (plus the unlisted reprise after “Fat Chance Hotel” that closes out Happy?) by Public Image Ltd, “Same Old Scene” by Roxy Music, “Driver’s Seat” by Sniff ‘n’ the Tears and “Cuad El Habib” by Yello – occupied five slots.

No, I began to “weight” appearances on some mixes more than others. Nearly 30 years and many iterations later, I can only guess at those first weights – but an appearance on the Boston Drive mixes equated to something like three slots, while appearing on the four “non-series” mixes something like two slots, as did appearance on particularly beloved mixes like Stuff and Such Vol. XXX, created in June 1992. It is possible appearance on “one-artist-only” mixes counted as <1 slots, but I doubt it – I had not yet reached that level of, umm, methodological sophistication.

Ranking Hypothesis #4: I like tracks included on designated mixes more than those only included on non-designated mixes.

Ranking Hypothesis #4a: How much more I like tracks included on designated mixes varies by designated mix.

At the time, ignoring 56.1% of the 1,166+ tracks I had recorded over the previous 11+ years made logistical sense. It also allowed me to ignore the distinction between “now/recently” and “of all time.” I continued to exclude them – with a partial exception I revisit in the next installment – for more than a decade.

It was thus a limited collection of 512 tracks from which I tabulated – using the “ranking hypotheses” listed above – the first installment of “The Berger 100;”[3] “The Berger 10” received its own page:

10. “Promised You a Miracle”

9. “Are ‘Friends’ Electric?” Tubeway Army[4]

8. “The Evening’s Young” Yello

7. “New Toy” Lene Lovich[5]

6. “Stay Hungry” Talking Heads

5. “Cuad El Habib”

4. “Entangled” Genesis

3. “Right Down the Line” Gerry Rafferty

2. “Driver’s Seat”

1. “Save Me/Reprise”

“Zamba” was #12 and “Same Old Scene” was #19.

Based on what I remember of that time, the “ranking hypotheses” – the earliest incarnation of a score-computing algorithm – worked very well, at least in terms of what I most listened to in the early 1990s. From the song rankings, I generated “The Berger Album 50/10” and the “The Berger Artist 100/10.” At least, I think I did – I dated the track listings but not those for album and artist. Not that it mattered, as Ultravox’s Quartet and Genesis, respectively, topped those lists throughout the 1990s.

We return to that decade in the next installment, in which mix-making protocols emerge, mixes proliferate and technology flummoxes me…before it makes me rejoice.

Until next time…please be safe and healthy – and if you not already done so, please get vaccinated against COVID-19!

[1] I prefer “track” to “song” because it encompasses the full range of “music-related things that can be recorded onto a mix” tape, CD or video.

[2] For the previous two years, an older Harritonite rode the same bus as me, and she regularly played a mix of 60s folk rock tunes. The side I recorded – Beatles, Rolling Stones, Chicago, Seals and Crofts, Moody Blues, David Bowie, John Lennon – was inspired by her mix.

[3] Brian Eno’s “Julie With” was #100.

[4] Incorrectly written as Are Friends Electric? Gary Numan

[5] Yes, that is Thomas Dolby in the video.

A Wicked Early Look At U.S. Senate Races in 2022

In two recent posts, I…

I conclude this “wicked early” look at the 2021-22 elections with an examination of Democrats’ prospects in the 34 United States Senate (“Senate”) elections currently scheduled for November 8, 2022 – when they will defend their slender majority: a 50-50 tie broken by Vice President Kamala Harris.[1] As with the 2021-22 gubernatorial elections, this is a view from 30,000 feet: what the “fundamentals” say about how likely it is Democrats continue to control the Senate in January 2023.

“Fundamentals” are the sum of…

  • State partisan lean (using my 3W-RDM)
  • Estimated incumbency advantage (see below)
  • National partisan lean, measured by the “generic ballot” question: some variation of “If the election for were held today, would you vote for the Democratic candidate, the Republican candidate, or some other candidate?”

To estimate incumbency advantage in Senate elections, I first calculated an “expected margin of victory” for the Democratic nominee in the 35 such elections held in 2020, the 35 held in 2018 and the 34 held in 2016. “Expected margin of victory” is state 3W-RDM plus the difference between the Democratic percentage and Republican percentage of all Senate votes cast in each election cycle:

2020 = D-1.9 percentage points (“points”)

2018 = D+9.9 points

2016 = D+0.9 points

For 2020 elections, I…

  • Used results from the Georgia Senate runoff elections
  • Used number of votes cast for all Republican candidates, for all Democratic candidates and for all third-party candidates in the Louisiana Senate “jungle primary”[2] held on Election Day
  • Treated Libertarian nominee Ricky Dale Harrington, Jr. as “Other” in the Arkansas Senate race, despite no Democratic nominee

For 2018 elections, I…

For 2016 elections, I…

After subtracting actual margin from “expected” margin, I calculated three averages of these differences within each election year:

  1. Races with Democratic incumbents
  2. Races with Republican incumbents
  3. Open-seat races

Within each biennial election cycle, then, incumbency advantage for Democratic Senate candidates is the first average minus the third average (D+17.0 in 2020, D+0.9 in 2018, D+6.5 in 2016), while Republican Senate incumbency advantage is the second average minus the third average (R+4.4, R+2.6, R+3.6). It is not clear why Republican Senate incumbents have a more stable – and lower overall – estimated advantage than their Democratic counterparts.

The estimated effect of incumbency for each party heading into 2022 is this calculation using all 104 elections: +5.5 points for Democrats and +3.3 points for Republicans. If an incumbent has served less than a full four-year term – for example, Democrats Mark Kelly and Raphael Warnock won special elections in 2020 while Democrat Alex Padilla was appointed to the Senate seat vacated by Harris, and will face reelection in 2022 – I multiply incumbency advantage by approximate percentage of term served.


Turning to the elections themselves, I mimic the 2021-22 gubernatorial elections post by exploring three different scenarios:

  • Democrats win national House vote by 3.5 points, as current polls suggest
  • Democrats and Republicans split national House vote, assuming a sharp break by undecideds toward the out party
  • Republicans win national House vote by 3.5 points, in line with recent elections

Election data – unless otherwise specified – come from Dave Leip’s invaluable website.

2022 Senate elections – 14 Democratic incumbents/open seats:

NameStateRun 20223W-RDMINCNat LeanTotalLast marginFirst elected/apptd
Brian SchatzHIYes29.05.53.538.051.3%2012
Patrick LeahyVTYes28.95.53.537.928.2%1974
Chris Van HollenMDYes26.25.53.535.225.2%2016
Alex PadillaCAYes24.91.83.530.22021
Chuck SchumerNYYes20.25.53.529.243.4%1998
Richard BlumenthalCTYes13.95.53.522.928.6%2010
Patty MurrayWAYes13.75.53.522.718.0%1992
Tammy DuckworthILYes13.35.53.522.315.1%2016
Ron WydenORYes10.15.53.519.123.3%1996
Michael BennetCOYes5.75.53.514.75.7%2009
Maggie HassanNHYes1.25.53.510.20.1%2016
Catherine Cortez-MastoNVYes-
Mark KellyAZYes-
Raphael WarnockGAYes-

After losing nine Senate seats in 2014, Democrats are only defending 14 seats in 2022, and they are on very favorable turf to do so. Not only did Democratic presidential nominee Joseph R. Biden Jr. win all of these states in 2020, their average partisan lean is D+12.4 – D+15.6 if you exclude trending-Democratic-but-still-Republican-leaning Arizona and Georgia. Moreover, not a single Democratic incumbent is retiring, reflecting confidence they will retain their majority.

Ten of these 14 incumbents – Senate President pro tempore Pat Leahy of Vermont, aiming for a ninth Senate term at the age of 82, and those in Hawaii, Maryland, California, New York, Connecticut, Washington, Illinois, Oregon and Colorado – are heavy favorites to win reelection; even if Republicans have a monster night in 2022 – winning the national House vote by an astonishing 7.0 points – and even Michael Bennet of Colorado would still be a 7-3 favorite.

That leaves four remotely vulnerable seats. Going in reverse order of retention likelihood, we find two swing-state incumbent Democrats seeking a second term: Maggie Hassan of New Hampshire and Catherine Cortez-Masto of Nevada. Even in a relatively pro-Republican environment, both would be moderately favored – likely winning by low single digits. The wild card in New Hampshire is popular Republican Governor Chris Sununu, who, despite being a safe bet to win reelection in 2022, is being heavily recruited to run against Hassan, herself a former governor. Hassan won by just 1,017 votes in 2016, defeating first-term Republican Kelly Ayotte. Sununu’s entry would make this race a pure toss-up – an epic battle of political heavyweights. In Nevada, meanwhile, former state Attorney General Adam Laxalt, who lost the 2018 gubernatorial election by just 4.1 points in a strong Democratic year, is reportedly gearing up to run. Laxalt, like Sununu, is the son of a former governor, making him a potentially formidable opponent.

History and “fundamentals,” finally, suggest Kelly and Warnock are moderate underdogs for reelection, especially if we only credit them with one-third the estimated incumbency advantage. Even in the unlikely event Democrats re-run their 2020 turnout, and their opponents are weakened – either distracted by the ongoing Arizona 2020 ballot “audit” or lacking election experience, like former National Football League and University of Georgia Herschel Walker – both men are toss-ups at best. That all said, the worst-case scenario is that they only have a 1-in-6 chance of winning – equivalent to a single number coming up on a fair die roll.

Bottom line: Democrats will lose between zero and four Senate seats, with one-to-three likeliest.

2022 Senate elections – 20 Republican incumbents/open seats:

NameStateRun 20223W-RDMINCNat LeanTotalLast marginFirst elected/apptd
Pat ToomeyPANo-
Ron JohnsonWI???-2.4-3.33.5-2.23.4%2010
Marco RubioFLYes-5.5-3.33.5-5.37.7%2010
Richard BurrNCNo-
Rob PortmanOHNo-
Chuck GrassleyIA???-9.8-3.33.5-9.624.4%1980
Lisa MurkowskiAKYes-15.8-3.33.5-15.632.7%2002
Tim ScottSCYes-15.9-3.33.5-15.723.6%2013
Roy BluntMONo-
Todd YoungINYes-19.6-3.33.5-19.49.7%2016
Jerry MoranKSYes-21.3-3.33.5-21.129.9%2010
John KennedyLAYes-22.3-3.33.5-22.121.3%2016
Mike LeeUTYes-27.6-3.33.5-27.441.1%2010
Richard ShelbyALNo-
John ThuneSDYes-29.6-3.33.5-29.443.7%2004
Rand PaulKYNo-30.3-3.33.5-30.114.6%2010
John BoozmanARYes-30.3-3.33.5-30.123.6%2010
Mike CrapoIDYes-34.8-3.33.5-34.638.4%1998
John HoevenNDYes-35.4-3.33.5-35.261.5%2010
James LankfordOKYes-37.8-3.33.5-37.643.2%2014

Like Democrats, Republicans are defending Senate seats on extremely favorable turf: the average partisan lean of these 20 states is R+20.2, but take away the swing/lean Republican states of Pennsylvania, Wisconsin, North Carolina and Florida, and it jumps to R+24.3. Which makes all the more puzzling why six Republican Senators have already announced plans not to seek reelection in 2022 – and Ron Johnson of Wisconsin and seven-termer Chuck Grassley of Iowa have not yet announced either way. It is a truism that an early wave of retirements by elected officials of one party presages an electoral walloping for that party – but in this case that is the Republicans, who only need to net one seat in what should be a good year for Republicans to win back the Senate.

For all that, though, I anticipate easy Republican wins in Oklahoma, North Dakota, Idaho, Arkansas, Kentucky, South Dakota, Utah, Louisiana, Kansas, Indiana and South Carolina. And as much drama as former president Donald J. Trump and Senate Minority Leader Mitch McConnell may cause in the Republican Senate primary in Alabama – where Democrat-turned-Republican Richard Shelby is retiring after six terms – the Republican nominee is still heavily favored to win. And we should probably add R+9.8 Iowa to this list, especially if Grassley runs for reelection.

Which brings us to two wild-card races: Lisa Murkowski’s reelection bid in Alaska and the open seat left by Roy Blunt’s retirement in Missouri. Murkowski – seeking her fourth full term in office – would normally be a shoo-in to win reelection in R+15.8 Alaska. However, Alaska will use a “jungle primary” for the first time in 2022 – and Murkowski, who has drawn the ire of Trump and his allies for her relatively moderate voting record, will face a stiff challenge from Republican Kelly Tshibaka, a former commissioner of the state Department of Administration. If Murkowski finishes in the top two, she would likely prevail in the runoff. But if the runoff is between, say, Tshibaka and Independent 2020 Senate nominee Al Gross – anything is possible.

In Missouri, meanwhile, the spanner in the works is former governor Eric Greitens, who resigned amid criminal clouds in June 2018 – then announced his Senate candidacy on March 22, 2021. Blunt only won reelection by 2.8 points in 2016 – a relatively neutral year – so despite Missouri being R+19.0, Greitens as the GOP nominee could make this race a toss-up. Possibly. Maybe.

On balance, however, Republicans are still more likely than not to win all 15 of these seats. And the “fundamentals” suggest they should also be favored to retain the seat being vacated by Rob Portman in R+9.8 Ohio, although the announcement by Ohio House Member Tim Ryan that he is running for the Democratic nomination decreases that probability to perhaps “only” 2-1.

Tim Ryan may be the only Democrat who could win an open Senate seat in Ohio in 2022.

Similar hope accrues to the recent announcement by Florida House Member Val Demings that she will seek the Democratic Senate nomination against two-term Republican Marco Rubio. Florida has drifted from being a true swing state to one 5.5 points more Republican than the nation as a whole, and Rubio will have the advantages of incumbency and (probably) a pro-Republican environment. Nonetheless, Demings has an expanding national profile after reportedly being on Biden’s vice-presidential short list and serving as a House Manager in Trump’s first impeachment trial. Thus, like Ryan in Ohio, a Demings nomination likely reduces the odds to “only” 2-1 against.

Could Val Demings defeat two-term Republican Marco Rubio in Florida in 2022?

This brings us, finally, to the three most winnable elections for Democrats according to the fundamentals: North Carolina, Wisconsin and Pennsylvania. Despite being 5.8 points more Republican than the nation as a whole, North Carolina – where three-term Senator Richard Burr is retiring – poses a strong pickup opportunity for Democrats (even in a bad year for them overall) because of what will likely be a messy and chaotic nomination process. While this could happen on the Democratic side, where former state Supreme Court Chief Justice Cheri Beasley is an early front-runner, I argue fear of losing the Senate makes them more likely to unite early around Beasley or some other candidate.

If Johnson seeks a third term in 2022, the combination of incumbency, Wisconsin’s Republican drift and a likely pro-Republican environment makes him something like a 3-1 favorite. But if he does not run, this race is much closer – though still far from a likely Democratic pickup. One problem is a lack of star power on the Democratic side – and no clear signal from Lieutenant Governor Mandela Barnes whether he will run for the Senate or seek reelection. Even then, this race would begin as a toss-up.

And, just as in 2020, control of the Senate may come down to Pennsylvania, where Pat Toomey is retiring after two terms – and two very narrow victories (2.0 points in 2010, 1.4 points in 2016). Effectively a toss-up, this is the Democrats’ best chance to pick up a state – especially if Lieutenant Governor John Fetterman wins the nomination, though state representative Malcolm Kenyatta – a young, openly-gay, black progressive – could drive base Democratic turnout.

Lieutenant Governor John Fetterman could join fellow Pennsylvania Democrat Bob Casey in the Senate in January 2023.

As could State Representative Malcolm Kenyatta

Bottom line: There is no obvious Republican loss, only good-to-solid opportunities in Pennsylvania, North Carolina and Wisconsin; less-good opportunities in Florida and Ohio and hard-to-decipher long shots in Alaska and Missouri. These races effectively come down to a battle between historic trends – which suggest Republicans should win all of these races – vs. Republican division/Democratic unity – giving Democrats a realistic shot at winning from one to five seats. This suggests a practical pick-up range for Democrats of zero-to-three seats.

Overall outlook. The fundamentals – a Republican-leaning map (median 2022 Senate election = R+6.5 state), a rough balance in incumbency and what should be a pro-Republican environment – say Democrats have an uphill battle to retain the Senate in 2022. Democrats have two seats more likely than not to flip – Arizona, Georgia – and two other vulnerable seats – Nevada and New Hampshire, while the best Democratic pickups opportunities – Pennsylvania, North Carolina, Wisconsin, Florida and Ohio – still lean Republican. That said, Democrats have no open seats, while Republicans have at least six, which historically signals Democratic strength. And Democratic candidate recruitment has strongly outclassed that of Republicans – who would apparently rather fight proxy Trump-McConnell battles than nominate strong candidates.

In short, anything from Democrats losing a net four seats to Democrats gaining six seats is broadly plausible…though a range of a net loss of three to a net gain of three is as narrow as I am willing to go at this point. Make of it what you will that the midpoint of this range is status quo—and continued Democratic control of the Senate.

Until next time…please be safe and healthy – and if you not already done so, please get vaccinated against COVID-19!

[1] Two Independents – Angus King of Maine and Bernie Sanders of Vermont – caucus with Democrats.

[2] All candidates run in same election regardless of party; if no candidate tops 50%, top two finishers advance to runoff

A Wicked Early Look At Governor’s Races in 2021 and 2022

In a recent post, I assessed it was fairly likely Republicans regain control of the United States House of Representatives (“House”) in 2022. In this post, I turn my attention to the two gubernatorial elections to be held in 2021 (New Jersey, Virginia) and the 36 gubernatorial elections to be held in 2022. My goal is to provide the view from 30,000 feet: what the “fundamentals” in each race reveal about the overall partisan landscape—and how likely it is Democrats cut into the current 27-23 Republican gubernatorial advantage.

“Fundamentals” are the sum of three values:

  1. State partisan lean, measured using my three-year-weighted relative Democratic margin (“3W-RDM”), or weighted[1] three-election average of the difference between a state’s Democratic (minus Republican) margin in a presidential election and the Democratic (minus Republican) margin in the total national vote in that election. For updated 3W-RDM following the 2020 presidential election, see here.
  2. Estimated incumbency advantage (incumbent office-holders tend to receive a higher percentage of the vote than an open-seat candidate of the same party).
  3. National partisan lean, measured by the “generic ballot” question (variations on “If the election for were held today, would you vote for the Democratic candidate, the Republican candidate, or some other candidate?”)


To estimate incumbency advantage in gubernatorial elections, I first calculated an “expected margin of victory” for the Democratic nominee in the 14 gubernatorial elections in 2019 (3[2]) and 2020 (11) and the 38 gubernatorial elections in 2017 (2) and 2018 (36); New Hampshire and Vermont elect governors every two years. I had previously estimated the effect of gubernatorial incumbency using two full cycles, but I decided to keep these calculations in line with those for Senate incumbency (last three cycles: 2016, 2018, 2020) as possible.

“Expected margin of victory” is state 3W-RDM plus the difference between the Democratic percentage and Republican percentage of all gubernatorial votes cast in each two-year period:

2019-20 = D-7.0 percentage points (“points”)

2017-18 = D+3.5 points

After subtracting actual margin from “expected” margin, I calculated three averages of these differences within each election year:

  1. Races with Democratic incumbents
  2. Races with Republican incumbents
  3. Open-seat races

Within each two-year election cycle, then, the effect of incumbency for Democratic candidates for governor is the first average minus the third average (D+23.3 in 2019-20, D+2.0 in 2017-18), while the Republican advantage is the second average minus the third average (R+10.9, R+17.3). And the estimated effect of incumbency for each party is this calculation using all 52 elections: +10.4 points for Democrats and +13.9 points for Republicans. If an incumbent has served less than a full four-year term –  as Rhode Island’s Democratic Lieutenant Governor Dan McKee has since Democratic Governor Gina Raimondo was sworn in as Secretary of Commerce – I multiply incumbency advantage by the percentage of term served, using an approximation of 50% for McKee.

The effect of incumbency for gubernatorial elections remains very strong: heading into the 2019-20 cycle it was D+5.7 and R+8.5 (albeit based on 104 elections over eight years).


We now turn to the elections themselves…and immediately face the problem of estimating partisan lean as of November 2022. In 2022 House elections post, I estimated that the current partisan lean is Dem+3.5, though with an average of 14.3% of the electorate undecided or leaning toward third-party candidates. If the vote for third-party candidates is 1.5%, and true undecideds split 2-1 Republican, partisan lean is GOP+0.7. Even that latter number may be wishful thinking for Democrats – Democrats (in governor’s races, anyway) led 3.5 points in 2018, the last time a newly-elected president – in this case Republican Donald J. Trump – faced a first midterm election.

To account for this uncertainty, we explore three different scenarios in 2019-20:

  • Democrats win national House vote by 3.5 points, as current polls suggest
  • Democrats and Republicans split national House vote, after a sharp break by undecideds toward the out party
  • Republicans win national House vote by 3.5 points, in line with recent elections

Election data – unless otherwise specified – come from Dave Leip’s invaluable website.

2021 gubernatorial elections (November 2):

NameStateRun 20213W-RDMINCNat LeanTotalLast marginFirst elected
Phil MurphyNJYes12.010.43.525.914.2%2017
Ralph NorthamVANo3.903.57.4N/AN/A

In New Jersey, Democratic Governor Phil Murphy seeks reelection against Republican Jack Ciattarelli. New Jersey’s strong Democratic lean and incumbency advantage suggest Murphy will cruise to reelection irrespective of the national political environment. Two publicly-available polls of this election, conducted in May 2021 by B/C level pollsters, show Murphy ahead by an average 50-31%, which aligns with a pro-Republican environment.

In Virginia, Democratic Governor Ralph Northam is limited to one term, so former Democratic Governor Terry McAuliffe will face Republican Glenn Youngkin. Virginia has been trending sharply Democratic, but without the advantages of “direct” incumbency (I am not sure how to assess “formers”), McAuliffe is only modestly favored to win back the job he left in 2018. So long as the environment is pro-Democrat or neutral, McAuliffe should prevail by 3-8 points…but if it shifts pro-Republican, this race becomes a pure toss-up. Two publicly-available polls of this election, conducted in June 2021 by B/C level pollsters, show McAuliffe ahead by an average 47-44%, which aligns with a neutral environment.

California Governor Gavin Newson will also face a recall election sometime in 2021. It is actually two elections. The first is a “yes” or “no” on Newsom. If “yes” prevails, Newsom remains governor. If “no” prevails, Newsom is booted and voters then choose from a slate of candidates excluding Newsom. Based solely on the recent improvement in California’s finances, I expect Newsom to survive the first vote relatively easily.

Bottom line: While Virginia could be close, Democrats are unlikely to lose any governor’s mansions in 2021.

2022 gubernatorial elections – 16 Democratic incumbents/open seats (November 8):

NameStateRun 20223W-RDMINCNat LeanTotalLast marginFirst elected
David IgeHINo29.00.03.532.5N/A
Gavin NewsomCA???24.910.43.538.823.9%2018
Andrew CuomoNY???
Dan McKeeRIYes16.65.23.525.32021
Ned LamontCTYes13.910.43.527.83.2%2018
J.B. PritzkerILYes13.310.43.526.215.7%2018
Kate BrownORYes10.10.03.513.6N/A
Michelle Lujan GrishamNMYes6.310.43.520.214.4%2018
Jared PolisCOYes5.710.43.519.610.6%2018
Janet MillsMEYes4.510.43.518.47.7%2018
Tim WalzMNYes1.810.43.515.711.4%2018
Steve SisolakNVYes-0.510.43.513.44.1%2018
Gretchen WhitmerMIYes-0.710.43.513.29.6%2018
Tom WolfPANo-
Tony EversWIYes-2.410.43.511.51.1%2018
Laura KellyKSYes-21.310.43.5-7.45.1%2018

With the exception of Kansas, these are either strong Democratic states, Democratic-leaning states or swing states, minimizing obvious losses; the average partisan lean is D+7.4, D+9.4 without Kansas. And even if Newsom loses the recall vote, Democrats are still favored to regain that governor’s mansion, along with those in Hawaii, New York, Connecticut and Illinois. One fly in the ointment in New York is whether embattled Governor Andrew Cuomo seeks a fourth term, though I expect the Democratic nominee to prevail easily if he does not.

Democratic incumbents are also favored in New Mexico, Colorado, Maine, Minnesota, Nevada and Michigan – as is the Democratic nominee in the seat being vacated by Oregon Governor Kate Brown. However, it is unlikely Mills, Sisolak and Whitmer will win by more than even their relatively modest margins in 2018. Regardless of the margins, though, Democrats are favored to retain 13 of the 16 governor’s mansion they will defend in 2022.

That leaves – in reverse order of retention likelihood – Wisconsin, Pennsylvania and Kansas. Tony Evers benefitted from a strong pro-Democratic environment and unpopular incumbent Scott Walker seeking a third term in 2018, yet still only won by 1.1 points in Republican-drifting Wisconsin. Incumbency makes him a narrow favorite to win again; the 4.5 points suggested by a pro-Republican environment feels about right. Tom Wolf is term-limited after defeating a Republican incumbent in 2014 by nearly 10 points, then winning reelection four years later by nearly 17 points. Pennsylvania, like Wisconsin, is drifting Republican, and it has a history – albeit broken by Wolf – of  flipping its governor’s mansion between the two parties every eight years. Thus, the fundamentals suggest Republicans are modest favorites to win it back in 2022 – unless Democratic front-runner Josh Shapiro, currently state Attorney General, secures the nomination against a divided and little-known Republican field. Finally, Laura Kelly – who won an open gubernatorial election in Kansas in 2018 against a very unpopular opponent in Kris Kobach, succeeding even more unpopular Republican governor Sam Brownback – is a solid underdog based on the fundamentals.

Pennsylvania Attorney General is likely Democrats’ best chance to retain the commonwealth’s governor’s mansion.

Bottom line: Democrats will likely lose one, and possibly two or three, of the 16 governor’s mansions they will defend in 2022.

2022 gubernatorial elections – 20 Republican incumbents/open seats (November 8):

NameStateRun 20223W-RDMINCNat LeanTotalLast marginFirst elected
Phil ScottVT???28.9-13.93.515.041.1%2016
Larry HoganMDNo26.20.03.529.7N/A
Charlie BakerMA???26.1-13.93.515.733.5%2014
Chris SununuNH???1.2-13.93.5-9.231.8%2016
Ron DeSantisFLYes-5.5-13.93.5-15.90.4%2018
Doug DuceyAZNo-
Brian KempGAYes-6.4-13.93.5-16.81.4%2018
Kim ReynoldsIAYes-9.8-13.93.5-20.28.7%2017
Mike DeWineOHYes-9.8-13.93.5-20.23.7%2018
Greg AbbottTXYes-12.0-13.93.5-22.413.0%2014
Mike DunleavyAKYes-15.8-13.93.5-26.27.0%2018
Henry McMasterSCYes-15.9-13.93.5-26.38.0%2017
Pete RickettsNENo-
Bill LeeTNYes-27.2-13.93.5-37.621.0%2018
Kay IveyALYes-29.2-13.93.5-39.619.1%2017
Kristi NoemSDYes-29.6-13.93.5-40.03.4%2018
Asa HutchinsonARNo-
Brad LittleIDYes-34.8-13.93.5-45.221.6%2018
Kevin StittOKYes-37.8-13.93.5-48.212.1%2018
Mark GordonWYYes-47.5-13.93.5-57.939.6%2018

Like Democrats, Republicans are defending governor’s mansion on very favorable turf -with four glaring exceptions in New England and Maryland; the average partisan lean of these 20 states is R+13.0, but take away the four swing/strong Democratic states, and it jumps to R+21.4.

And while many of the margins in the table above seem ludicrous – I suspect incumbency advantage dissipates when your party dominates – Republican governors are heavy favorites to win reelection in Tennessee, Alabama, South Dakota, Idaho and Wyoming, while holding onto open governor’s mansions in Nebraska and Arkansas. Henry McMaster in South Carolina, Mike Dunleavy in Alaska and Kim Reynolds in Iowa should also win reelection, perhaps hitting low double-digits in what could be a good Republican year.

Assuming Greg Abbott seeks a third term as governor of Texas, he should win reelection – if he can survive primary challenges from his right. It is unlikely the collapse of Texas’ power grid earlier this year will impact voting. The same is true of Ohio Governor Mike DeWine, who only defeated Democrat Richard Cordray by 3.7 points in 2018, in a state nearly 10 points more Republican than the nation as a whole.

Similarly, Georgia Governor Brian Kemp defeated Democrat Stacey Abrams by just 1.4 points in 2018, albeit in a very Democratic year. IF she runs again, as she has strongly hinted, and if Kemp himself survives a primary challenge, this is a sleeper pickup opportunity for Democrats.

Will former Georgia House minority leader Stacey Abrams run for governor again in 2022?

Another sleeper pickup opportunity is Arizona, where Governor Doug Ducey is term-limited. Like Georgia, Arizona both voted for Democrat Joseph R. Biden, Jr. in 2020 and in the space of two elections went from two Republican United States Senators (“Senators”) to Democratic ones. Just as Ohio, Iowa and Florida are becoming more Republican, Arizona and Georgia are becoming more Democratic – or, rather, less Republican. A Democratic nominee like state Secretary of State Katie Hobbs would still be at best a slight underdog, as would be Abrams.

Arizona Secretary of State Katie Hobbs could be the Democratic nominee for governor in Arizona in 2022.

Speaking of Florida, Ron DeSantis – who only won the open governor’s mansion in 2018 by 0.4 points – is clearly using a solid reelection in 2022 to run for the Republican presidential nomination in 2024. And, despite coming very close in recent elections, a Democrat has not been elected governor of Florida since 1994. As a Republican running for reelection in a state 5.5 points more Republican than the nation as whole in what could be a good Republican year, DeSantis should win easily. And yet…those recent narrow margins – especially the 1.1-point margin in the strong Republican year of 2014 give me pause, and suggest this should be in the sleeper column with Arizona and Georgia.

That leaves Maryland and the three New England states, where moderate or center-right Republicans continue to win gubernatorial elections. Maryland, where Larry Hogan is term-limited, is by far the strongest opportunity for Democrats to win back a governor’s mansion in 2022: even in a strong Republican year, a generic Democratic nominee is still heavily favored. As for the three New England states, the only question is whether Charlie Baker in Massachusetts, Chris Sununu in New Hampshire and/or Phil Scott in Vermont seek reelection. Sununu is being heavily recruited to run against Democratic Senator Maggie Hassan; if he runs, the gubernatorial election becomes a pure toss-up. Moreover, if Baker and/or Scott – whose average margin of victory in their previous elections was 37.3 points – do not seek reelection, the Democratic nominee would be heavily favored to win. If either does run, though, he would be the favorite – though Massachusetts Attorney General Maura Healey could make things very interesting for Baker.

Will state Attorney General Maura Healey, a Democrat, be the next governor of Massachusetts?

Bottom line: On paper, Republicans are very likely to lose one governor’s mansion (Maryland), could easily lose another one-to-three (Massachusetts, New Hampshire, Vermont) under the right circumstances, and could see another three become more competitive (Arizona, Florida, Georgia). Outside of Maryland, Republican success hinges upon whether incumbents seek reelection in New England, how right-wing primary challenges fare and just how Republican 2022 proves to be – if it is at all.

Overall outlook. The bad news for Democrats is that unless they maintain high levels of voting enthusiasm, Republicans could do very well in 2022 – though not in 2021. The good news for Democrats is that they have more overall opportunities. Based solely on the fundamentals, they have at most three vulnerable seats – they are a heavy underdog in Kansas, a slight underdog in Pennsylvania, and only a slight favorite in Wisconsin – while Republicans could have anywhere from one to seven (though keep an eye on Texas). Realistically, however, Democrats are likely to lose one or two of their governor’s mansions while netting one-to-three, putting them somewhere between a net loss of one and a net gain of two. All things considered, though, what amounts to a draw could be considered a win for Democrats in 2022.

Until next time…please be safe and healthy – and if you not already done so, please get vaccinated against COVID-19!

[1] Most recent election weighted “3,” 2nd-most recent “2,” 3rd-most recent “1”

[2] For the 2019 Louisiana gubernatorial election, I used data from the runoff election held November 16, 2019.

How Likely Is Republican Control of the U.S. House In 2022?

On March 31, 2021, nearly five months after Election Day 2020, Democrat Rita Hart finally conceded to Republican Mariannette Miller-Meeks in Iowa’s 2nd Congressional District (“CD”), dropping her challenge to her six-vote loss. This was a net win for the Republicans, as United States House of Representatives (“House”) Member Dave Loebsack, a Democrat, had not sought reelection to an 8th term.

Overall, Democrats lost a net of 13 House seats relative to Election Day 2018, going from a 235-200 majority to a 222-213 majority. This double-digit seat loss occurred despite Democrats winning a 3.1-percentage-point (“point”) majority in all votes cast for U.S. House – 50.8 to 47.7%; notably, this was a 5.5-point decline from 2018

Returning to 2020, though, political observers were shocked at the number of House seats Democrats lost, especially because they won the presidency and netted three United States seats, giving them a 50-50 tie broken by Democratic Vice President Kamala Harris. Anyone following this website since at least June 2017, however, should not have been at all surprised.

Figure 1, using data from the 26 House elections from 1970 (vs. 1968) to 2020 (vs. 2018), shows that change in national House vote percentage accounts for an astonishing (for social science) 83% of the variance in House seat gain/loss. Reassuringly, the ordinary least squares (“OLS”) regression fitted to the data crosses the y-axis very close to the origin: a 0-point change in vote percentage essentially equates to no change in seats, which makes intuitive sense.

Figure 1: Change in % Democratic of Total House Vote vs. Net Change in Democratic House Seats, 1970-2020

Removing the 2018-20 data point from the OLS regression, suggests a drop in the Democratic share of the total House vote of 5.5 points yields an estimated loss of…

Dem Seat Change = 3.21 * (-5.5) – 1.19 = -18.8

Based on election data from the previous 50 years, Democrats should have lost 19 seats I 2020, with a 95% confidence interval (“CI”) of 18-20 seats. Yes, Democratic leaders expected to pick up seats, but history presents a compelling counter-narrative: Democrats were extremely lucky to maintain their House majority, however slender.


But what about 2022, the first midterm election of the Administration of Democrat Joseph R. Biden, Jr., when Republicans only need to net five seats to regain the majority? Figure 2 shows how poorly the first House elections after a new president is elected go for that president’s party – and getting worse over time. Granted, there have only been nine such elections starting with the 1962 midterms of Democratic President John F. Kennedy. Still, excepting the first post-9/11 midterm election – when Republicans under President George W. Bush in 2022 gained eight House seats – the party controlling the White House has lost an average of 53 House seats in the last 30 years. The overall average since 1962 is a net loss of 24 House seats…which would give Republicans a 237-198 edge going into 2023; the median is only -15, which would give Republicans a 228-207 edge.

Figure 2: 1st Midterm Seat Loss by Newly-Elected White House Party, 1962-2018

Based on the OLS regression equation in Figure 2, meanwhile, Democrats would be expected to lose 47 seats in 2022, dropping them to a 175-260 minority. However, because there are only nine data points, there is a great deal of “wobble” in this estimate – the 95% CI is a nonsensical -4,002 to +3,909.

In other words, all we really know from these nine 1st-term midterm elections is that anything between a loss of 61 seats and a gain of eight seats is historically plausible, with something in the 15-25 loss range most plausible.



This brings us full circle back to my model…and the question of what the margin in the total House vote will be in 2022. A little algebra reveals that to maintain a slender 218-217 House majority going into 2023, Democrats need to win the total House vote by at least 2.1 points – and likely at least 3.1 points (no net change) to account for a +/-1 seat 95% CI.

For comparison, political scientist Alan Abramowitz just published a model of net Democratic House seat gain. The two independent variables are 1) national House vote total (as opposed to change from previous election) and 2) number of House seats currently occupied by the president’s party. Those two variables account for 83% of the variance, as much as my single independent variable does. That said, Abramowitz’s model uses data from 1946-2018, giving him an additional 11 data points: 37 vs. 26. According to this model, Democrats would need to win the national House vote by about 6 points to maintain control, which equates to an increase of about 2.9 points from 2020.

Neither model is wrong, per sé. We are both trying to model a relatively infrequent occurrence while zeroing in a very precise outcome – the difference of a few seats – leading to the magnification of what really are small differences in outcomes. A 3-point lead and a 6-point lead are relatively close, and we agree Democrats have an uphill battle to retain the House, even without taking into account changes in CD lines due to reapportionment following the 2020 Census.

Technically, Abramowitz models Democratic lead on the “generic ballot question,” a variation of the poll question “If the election in your CD were held today, would you vote for the Democrat, the Republican, or some other party – or not vote at all.”

I recently compiled the results of the 20 generic ballot polls released publicly in 2021 then used the following steps to calculate a weighted-adjusted polling average (“WAPA”):

  1. Adjust raw margin – Democratic % minus Republican % – for “pollster bias,” as calculated for Pollster Ratings. Essentially, this is how much the pollster missed the final margin, on average, in recent polls of the same race.
  2. Average adjusted margins by how far the poll was conducted from Election Day 2022 – using midpoint of poll field dates – and pollster quality.
    • Time weight: (677 – days to Election Day)/677
    • Pollster weight: Numeric value of letter grade (A+ = 4.3, A = 4.0, etc.) divided by 4.3.

For example, the most recent generic ballot poll was conducted by Quinnipiac University from May 18 to May 24, 2021; it shows Democrats leading 50-41%. Quinnipiac has a historic Democratic skew of 0.5 points, meaning that 9.0-point lead is effectively an 8.5-point lead. The field midpoint was May 21, 2021, or 536 days until November 8, 2022, giving the poll a time weight of (677-536)/677 = 0.208. Their A- rating equates to a 3.7/4.3 = 0.860 pollster weight. Overall, this adjusted Democratic lead of 8.5 points has a weight of 0.208 * 0.860 = 0.179 – which, while not especially high, is highest among the 20 polls.

The raw average of these 20 polls is Dem+4.1, though mean Democratic bias of 0.5 points means this average is effectively Dem+3.6. Two polls without a pollster rating/bias were clearly conducted on behalf of Republicans, one in early April by PEM Management Corporation and one in late February by the National Republican Senatorial Committee. I traditionally used 1.5 as the bias for such polls, but I chose to match the most Democratic leaning poll in the 20: 3.5 for RMG Research. I also somewhat arbitrarily assigned these polls a letter grade of C; overall average was B-.

The initial WAPA I calculated was Dem+3.9, though it treated the multiple polls from RMG Research (2), McLaughlin & Associates (3) and Echelon Insights (4) as statistically-independent, even though polls conducted by the same firm are likely related. I thus calculated a second WAPA, which was the pollster-quality-weighted average of the bias-adjusted, time-weighted WAPA for each pollster. Basically, it is WAPA adjusted for pollster. This value was Dem+3.1

Averaging these WAPAs yields my best estimate of the November 2022 generic ballot as of June 2021: Dem+3.5. This is astonishing because it means support for Democratic House candidates increased 0.4 points since November – when, historically, support for the “out party” should be increasing.

If, in fact, Democrats win the national House vote by 3.5 points in November 2022, I estimate they would gain 0.3 seats, albeit with a 95% CI of -4.1 to +4.8 – which means they would have approximately a 94% chance of retaining the House. The Abramowitz model, however, estimates a 9-seat loss in this scenario – losing the House in the process.

It is not at all clear, of course, how well a 3.5-point lead in June 2021 translates to actual voting in November 2022. My poli-sci-sense suggests this lead – fairly robust since January, mind you – will slowly fade over the next year-plus due to a traditional complacency on the part of infrequent voters, in this case Democrats who may reason that as long as Biden is president, they do not need to vote in 2022, and a renewed enthusiasm on the part of out-party voters. This differential in voting enthusiasm, I suspect, is what leads to lopsided out-party victories in midterm elections.

Another reason for extreme caution is that this Dem+3.5 margin equates to Democrats 44.6%, Republicans 41.1%, Other/Undecided 14.3%. If Other earns the same 1.5% it did in 2020, that leaves fully 12.8% of voters up for grabs. It is not unreasonable – given voting enthusiasm differences – they split 2-1 for Republicans: roughly 8.5-4.3. This would actually give Republicans a national House vote lead of 49.6-48.9% on Election Day, a Democratic vote decrease of 3.8 points and a loss of 13 House seats (+/-2), very close to the recent median of -15.

Of course, if those undecided voters split evenly, Democrats are back to +3.5; a break toward Democrats seems extremely unlikely. So, let’s split the difference: a Democratic national House vote lead of 1.4 points, which equates to a loss of 6.3 seats, albeit with a 3.1 to 9.5 95% CI, giving Democrats something like a 15-20% chance of retaining their majority. This is as close to a “forecast” as I am willing to come in June 2021.

The bottom line is this: Republicans are favored to win back the House in 2022, though whether extremely narrowly or lopsidedly is far from clear. Historic trends in 1st midterm elections – of which there are only nine since 1960 – suggest Democrats could lose anywhere from 15 to 47 seats. Models with more data points – though still only 26-37 – suggest the shift is likely to be much smaller, anywhere from Democrats essentially holding serve – even netting a seat – to a Republican seat gain in the low double-digits.

Basically, keep an eye on the generic ballot numbers – if they stay close to Democrats ahead 3-4 points, they could be on the verge of defying decades of recent political history. If it drops closer to even, or Republicans pull slightly ahead – it will be a long night for House Democrats on Election Day 2022.

Until next time…please stay safe and healthy – and if you have not already been vaccinated against COVID-19, please do so!

Finding The Worst Character In Neo-Noir: And The Winner Is…

At 2:44 am on May 8, 2021, I tweeted the following using the handle @drnoir33:

Finding the worst character in #neonoir begins with this Corrupt Power matchup:

Noah Cross from CHINATOWN

Harry Angel from ANGEL HEART

#filmnoir #cinema #film

Attached to the tweet was a poll allowing a Twitter user to choose either Cross or Angel.

Realizing voting was a bit sluggish, at 6:21 pm I tweeted…

            For context:

…adding a link to this post and attaching this:

After 36 hours, the poll ended. Cross had beaten Angel, 72% to 28%, with 13 votes for Cross and five for Angel. Had I voted (see below), it would have been 14-5.


Five days earlier, I published this post explaining the origins of my “search for the worst character in neo-noir.” Once voting began, a friend who tweets under the handle @disquiet sought clarification for “worst character,” wondering if it meant “poorly-written” as opposed to “villainous.” I assured him it was the latter, though I had hesitated to use the word because some bad characters like Angel are the nominal protagonists of their film.

Over the next two posts – the second of which is here – I described my character selection process, delineated four broad categories (Corrupt Power, Crime Boss, Cunning Manipulator, Psychotic Loner/Hired Assassin), explained how I “seeded” characters within each category, and reduced 64 characters to 16.

Shortly after publishing the third post, I tweeted the first poll, pitting Cross against Angel. Once voting ended, I tweeted the next Corrupt Power match-up: Dr. Hannibal Lecter vs. Dudley Smith.

While the first matchup was live, a fellow film enthusiast had opined that Smith, the corrupt Los Angeles homicide Captain from L.A. Confidential, was his choice to win the competition. I am now free to say he was my choice as well…but I had not realized that you cannot vote in your own Twitter poll. And my wife Nell voted for Lecter, arguing cannibalism trumps even Smith’s level of corruption.

It did not occur to me until after this match to create a different Twitter account solely to vote in my polls. Had I done so earlier, the Lecter-Smith matchup would have ended in an 8-8 tie. However, the public-facing results showed Lecter winning 8-7.

Which meant I faced a conundrum.

Do I announce my vote for Smith, then go to the tiebreaker – one I had not yet explained? Or do I leave the results as they were? My fear was that Smith would win the tiebreaker, and it would appear I had put my thumb on the scale – leading potential voters to deem the process “rigged.”

The tiebreaker itself was simple: generating a random number on my iPhone. If the number was below 0.500, the lower seed – in this case Smith (2) – advanced. If the number was above 0.500, the higher seed advanced. If the number was 0.500 exactly, I generated a new random number.

Mostly as an experiment, I mentally voted for Smith – then pushed “Rand”…and saw “0.972,” meaning Lecter advanced. As much as I wanted Smith to win, I was relieved. I tweeted none of this, opting simply to move on to the first Crime Boss matchup: Frank Booth vs. The Joker. For the purposes of this post, though, Lecter won 9-8.


The remaining six Not-So-Sweet Sixteen matchups went smoothly.

Crime Boss

Booth beat Joker 77% to 23% (10-3).

In my first vote as “NeoNoirLover30” (@NLover30), I selected The Joker.

Keyser Soze beat Marsellus Wallace 75% to 25% (12-4).

I voted for Soze, as did another Twitter friend, who noted that as bad as Wallace was, Soze operated at an entirely different level of evil.

Cunning Manipulator

Leonard Shelby beat Catherine (Black Widow) 62% to 38% (8-5).

Tom Ripley beat Catherine Tramell 73% to 27% (11-4).

I voted for Shelby and Tramell. I was mildly surprised neither woman advanced to the Less-Than-Great Eight.

Psychotic Loner/Hired Assassin

Vincent (Collateral) beat Kevin (Sin City) 74% to 26% (17-6)

So much for cannibalism.

Anton Chigurh beat John Doe 53% to 47% (10-9)

I voted for Kevin and Doe, meaning I was on the winning side of only three of the first eight matchups; this pattern would continue, for better or for worse. A total of 135 votes were cast in these eight match-ups, an average of only 17 votes per matchup, a somewhat embarrassing number. The votes were divvied up thus:

Character1st Round2nd Round3rd Round4th RoundTotal
Harry Angel5   5
Frank Booth10    
Catherine5   5
Anton Chigurh10    
Noah Cross14    
John Doe9   9
The Joker3   3
Kevin6   6
Hannibal Lecter9    
Tom Ripley11    
Leonard Shelby8    
Dudley Smith8   8
Keyser Soze12    
Catherine Tramell4   4
Marsellus Wallace4   4
TOTAL135   135


With the Less-Than-Great Eight set, I began to “market” these matchups more aggressively, retweeting multiple exhortations to vote and to an expanding set of fellow film enthusiasts, and using photographs to remind potential voters of the characters.

It seemed to work, as the first match-up garnered 46 votes:

Corrupt Power: Cross beats Lecter 72% to 28% (33-13).

I was surprised Cross won so easily, even though I voted for Cross. One reason lies in what one voter argued: The Silence of the Lambs is not actually neo-noir, because Clarice (misspelled in the tweet) Starling was not a flawed protagonist. That latter point is debatable – I could counter Starling’s rookie “irrational exuberance” nearly gets her killed by Buffalo Bill – but instead pointed out the agnosticism of my selection method. Eight publicly-available lists included Silence, so it passes muster.

Crime Boss: Booth beats Soze 58% to 42% (15-11).

This was a very tight vote – tied at 11 late – until Booth finally pulled away; I voted for Soze. One voter simply noted how terrifying she recalled Booth being; having recently rewatched Blue Velvet, it is difficult to argue with her. Moreover, so much of what we think we know about Soze comes from one of most unreliable narrators in cinema.

Cunning Manipulator: Ripley beats Shelby 71% to 29% (24-10).

After voting for Shelby – who literally uses his anterograde amnesia to manipulate himself into becoming a serial killer – I was shocked how easily the talented Mr. Ripley won. But, Nell – who voted for Ripley – spoke for the majority when articulating how truly despicable she thinks Ripley is.

These kinds of surprises were part of what made this process so much fun.

Psychotic Loner/Hired Assassin: Chigurh beats Vincent 83% to 17% (24-5).

Once again, the lopsided vote surprised me. One reason may be that as much as I tried to make clear “Vincent” was the Tom Cruise character in Collateral, at least one voter thought it was “Vincent Vega” from Pulp Fiction; he was incredulous he had made it this deep into the voting. I myself voted for Vincent because while Chigurh occasionally uses a coin flip to spare his victims, Vincent never spared anyone.

Overall, 135 votes were cast in the second round, meaning the average vote exactly doubled to a somewhat-less-humiliating 34 per matchup.

Character1st Round2nd Round3rd Round4th RoundTotal
Harry Angel5   5
Frank Booth1015  25
Catherine5   5
Anton Chigurh1024  34
Noah Cross1433  47
John Doe9   9
The Joker3   3
Kevin6   6
Hannibal Lecter913  22
Tom Ripley1124  35
Leonard Shelby810  18
Dudley Smith8   8
Keyser Soze1211  23
Catherine Tramell4   4
Vincent175  22
Marsellus Wallace4   4
TOTAL135135  270

And with that, the Villainy Four was set.


Despite lacking gender diversity, the Villainy Four covered a wide range of times and places: one character each from the 1970s (Cross), 1980s (Booth), 1990s (Ripley) and 2000s (Chigurh). Moreover, Cross operated in 1930s Los Angeles, Ripley in 1950s Italy, Chigurh in 1980 New Mexico, and Booth in an all-American town called Lumberton in a 1980s that felt like the 1950s.

To shake things up – and because the original category quadrants were simply arranged counter-clockwise alphabetically – I used initial seeds to determine the third-round matchups. Thus, Corrupt Power faced off against Cunning Manipulator, while Crime Boss faced off against Psychotic Loner/Hired Assassin.

Cross beats Ripley 89% to 11% (17-2).

This was an absolute beat-down, which would have been worse had one voter not admitted she voted for Ripley just to be contrarian. Meanwhile, @disquiet now bluntly stated his assumption Cross would win the entire competition. I demurred, anticipating a barn-burner championship vote.

Chigurh ties Booth 50%-50% (15-15); higher-seeded Chigurh wins tie-breaker (0.925)

This was a genuinely tough choice – for me and for everyone. I ultimately voted for Chigurh solely because Chigurh appears to survive at the end of No Country For Old Men, while Booth is killed in a shootout.

I was only mildly upset 50 (counting the tie-breaker) votes – an average of 25 per matchup – were cast in the third round; it may not have helped that, after seeing voting plummet on a previous Sunday, I waited until after the three-day Memorial Day weekend to post the first third-round matchup. After 14 matchups and 320 total votes, it all came down to Chinatown and No Country For Old Men, the two post-1966 films most often cited as film noir by my Opportunity-Adjusted POINTS metric.


Character1st Round2nd Round3rd Round4th RoundTotal
Harry Angel5   5
Frank Booth101515 40
Catherine5   5
Anton Chigurh102416 50
Noah Cross143317 64
John Doe9   9
The Joker3   3
Kevin6   6
Hannibal Lecter913  22
Tom Ripley11242 37
Leonard Shelby810  18
Dudley Smith8   8
Keyser Soze1211  23
Catherine Tramell4   4
Vincent175  22
Marsellus Wallace4   4
TOTAL13513550 320


At 3:02 pm EST on June 7, 2021, I tweeted:

After 16 possibilities and 14 pairings…

…the search for the worst character in #neonoir concludes with this epic matchup!

Who will be crowned?

You decide!

#FilmNoir #film #FilmTwitter #Cinema #movies #villains #Chinatown #NoCountryForOldMen

By now, a small pool of Twitter users anticipated these matchups, and voting was brisk early. Nonetheless, as had happened in previous matchups, nearly all of the votes were cast in the first eight hours of voting – little changed over the final 28 hours. Still, a tournament-high 53 votes were cast.

I found this my toughest vote by far, but I finally voted for Chigurh, reasoning his willingness to kill people – many people – himself and his almost-robotic persistence made him worse than Cross, who – other than raping his own daughter and trying to kidnap their daughter – did almost no dirty work himself. It was not necessarily a good argument, but there it is.

Perhaps not surprisingly, I was again on the wrong side (and @disquiet nailed it):

Cross beats Chigurh 66% to 34% (35-18).

Character1st Round2nd Round3rd Round4th RoundTotal
Harry Angel5   5
Frank Booth101515 40
Catherine5   5
Anton Chigurh1024161868
Noah Cross1433173498
John Doe9   9
The Joker3   3
Kevin6   6
Hannibal Lecter913  22
Tom Ripley11242 37
Leonard Shelby810  18
Dudley Smith8   8
Keyser Soze1211  23
Catherine Tramell4   4
Vincent175  22
Marsellus Wallace4   4

Cross was the heavy favorite going into the voting phase of the competition (and perhaps from the very beginning): for one thing, he was the only 1-seed to make it to the Not-So-Sweet Sixteen. Moreover, Chinatown is the only film not released between 1940 and 1959 to make the overall Top 100 by film noir POINTS; its 20 LISTS and 32.0 POINTS rank it #81. For the primary villain of THE neo-noir film to be named “Worst Character in Neo-Noir” suggest the wisdom of the crowd worked brilliantly this time.

Still, I was genuinely surprised how easily Cross won. He won his four matchups by an average of 3-1. Overall, he earned 26% of the 373 total votes cast – had he won his four votes 13-12 (using the average number of votes over 15 matchups), he would have won only 14% of total votes cast. Chigurh was 2nd at 18%, with Booth and Ripley garnering 11% and 10%, respectively. In total, the Villainy Four earned nearly two-thirds of all votes cast.


Having been inspired to create this tournament by the Noir Alley March Badness competition on Twitter – which Phyllis Dietrichson of Double Indemnity won handily – I decided to hold a bonus vote: Cross vs. Dietrichson to crown the “Worst Character in ALL of #NOIR.”

The vote went live at 5 pm EST on June 9; I only kept it open 24 hours. Almost immediately, a mutual-Twitter follow woman tweeted: “This is my personal Sophie’s choice.” I suspect many votes were effectively mental coin flips. After my vote for Dietrichson – reasoning her doing all she did as a woman in 1944 gave her the villainy edge AND she has survived as a renowned villainess for more than 75 years – she took an early lead. But then Cross began to pull away, and within a few hours it was clear who was going to win. I still urged people to vote, and 41 did (42 if Nell – a Cross voter – had not had Twitter login issues), but when it was over:

Cross beats Dietrichson 69% to 31% (29-13)

That is what you call dominance – and it fits the man who embodies powerful white entitlement at its – he acquires everything he wants and expects to do so. Moreover, after I tweeted “You called it, @disquiet,” he responded, “This result gives me faith in humanity,” to which I responded, “It gives me faith in aggregation.”

Of course, we both simply could have said: forget it, Twitter, it’s Chinatown.

Until next time…please stay safe and healthy…and get vaccinated against COVID-19 if you have not already done so!

Measuring Film Quality: Revisiting “Guilty Pleasures”

In February 2019, I posed a deceptively simple question:

What makes a pleasure “guilty?”

To answer this question, I focused on films, specifically those I had seen multiple times. I gathered publicly-available data on these movies in order to assess how these films were regarded by both critics and fans. At that time, there were 557 such films, though I excluded 23 Charlie Chan films, having discussed them in a previous post. With these data, I generated a “perceived quality” (“PQ”) score. By comparing how much I liked a film to its PQ, I compiled a list of 11 “guilty pleasure” films, those I love despite their relatively poor reputations.

Recently, I updated that list, both adding films I forgot to include the first time and watching others for a second time. I also updated all data from the Internet Movie Database (“IMDb”) and the online movie rating site Rotten Tomatoes (“RT”), learning the latter no longer provides a count of the number of RT users – as opposed to “critics” – choosing to rate a film. Instead, they provide a characterization such as “10,000+ ratings;” I removed it from the database.

The goal was less to reexamine guilty pleasures – that list barely changed – than it was to examine the process itself.


On June 5, 2021 – OK, early on the morning of June 6 – I re-watched the 1996 Ted Demme film Beautiful Girls. In so doing, I increased the number of films in the database to 638.

Table 1 summarizes 10 of the remaining 13 variables, excluding title, category (described below) and date first shown in the United States. “Maltin” is the number of stars (BOMB=0) assigned by film critic Leonard Maltin in either his 2003 or 2008 Leonard Maltin’s Movie and Video Guide.[1]

Table 1: Summary statistics for Film Ratings Measures

MeasureNMean (SD*)MedianMinimumMaximum
Year of Release6381973.3 (22.6)198119202019
Length (mins.)638102.5 (19.5)101.048  220
IMDb Score6387.1 (0.7)
IMDb Raters638104,186.7 (239,116)22,887.51692,359,960
Tomatometer57976.3 (21.2)830100
RT Critic Average5797.0 (1.3)
RT Critics63848.8 (60.4)340541
RT Audience Score63371.7 (17.8)762096
RT Audience Average6333.8 (0.4)
Maltin Stars6152.8 (0.7)304

*SD=standard deviation, a measure of how tightly packed values are around the mean: the smaller the value, the tighter the packing. In a normal distribution, 68% of values are within 1 SD, 95% are within 2 SD and 99% are within 3 SD.

Compared to the original 557 films, this set of 638 films:

  • Were released one year earlier (mean and median) – though maximum increased to 2019 (Avengers: Endgame) and SD by 1.7.
  • Were 0.9 minutes shorter, with minimum now 48 minutes (Sherlock Jr.), increasing SD 1.7.
  • Had basically the same average/median IMDb score – a weighted average of user 0-10 ratings – and range, 4.2 (The Adventures of Rocky & Bullwinkle) to 9.0 (12 Angry Men, The Dark Knight).
  • Saw a mean increase of ~25,000 IMDb raters, with median increase of ~3,800; SD increased ~55,000, echoing a wider range of 169 (Southside 1-1000) to 2,359,960 (The Dark Knight), reflecting passage of time and addition of recent films with >1 million IMDb raters: Shutter Island, The Prestige, The Avengers and Inception.
  • Saw average, median and SD of Tomatometer – percentage of RT critics deeming a film “fresh” – drop slightly, with range still 0 (Once Upon a Crime…) to 100 (n=44).
  • Saw no change in average, median and SD of RT critic rating – 0-10 scale – while range widened: 2.1 (Hexed) to 9.8 (Sherlock Jr.).
  • Saw average, median and SD of number of RT critics increase by 8.8, 4.0 and 17.9, respectively, reflecting widened range of 0 (Charlie Chan at the Race Track, Murder in the Big House) to 541 (Avengers: Endgame)
  • Brought no change to average, median and SD for RT Audience Score – Tomatometer for fans – while range increased slightly: 17 (Street of Chance) to 97 (12 Angry Men).
  • Saw average and median Audience Rating – 0-5 scale – increase by 0.3: the distribution simply shifted to the right, with same SD and new range of 2.3 (Rocky & Bullwinkle, The Opposite Sex and How to Live With Them) to 4.6 (n=5).
  • Brought no change to Maltin statistics – though BOMB films increased from four to five with addition of Yellowbeard.
  • Basically, while the films I added were slightly older and shorter on average, and raters increased with time, the distribution of ratings did not materially change with the addition of 81 movies and updated data – except RT Audience Ratings, which increased for most movies.

Put another way, the “median” film I have seen multiple times remains a good-but-not-spectacular film like the 1980 comedy 9 to 5, which is 109 minutes long, has an IMDb score of 6.9 from just over 29,000 rates, a Tomatometer of 83 with an average 7.0 rating from 41 critics, a RT Audience Score of 83 with an average 3.8 rating, and 3.5 stars from Maltin. Half of the films I have seen multiple times are better-rated overall than 9 to 5 and half are worse-rated.

But here we run into the problem I sidestepped in the original post – how to create a single quality measure for all 638 films, when I only have complete data for 561 (87.9%) of them, a decrease of 4.6 percentage points. The biggest problem is lackof Tomatometer and RT Critic Rating data for 59 films with fewer than five critics.


Missing data is perhaps the bête noire of data analysts. There are only two solutions: 1) only use cases with complete data or 2) use statistical methods to estimate missing cases. The first solution is reasonable if you are not trying to generate a single value for every case and/or data are missing at random. Clearly this solution will not work here. Not only do we want a single value for each film – and only using the five variables (year, runtime, IMDb score, number IMDb raters, number RT critics) with no missing data feels inadequate – there is a clear pattern to the missing data. For example, 17 of the 23 films lacking Matlin stars were released after 2008. Conversely, the five films lacking RT audience data were released between 1936 and 1952, while the 59 films missing RT critic data have release year and runtime of 1947 and 80.6 minutes, respectively, with all but four released between 1931 and 1957.

This leaves estimation. A straightforward way is to use ordinary least squares regression (“OLS”), analogous to Y = Slope*X + “Where line crosses Y-axis”, or y = m*x + b, the formula we learned when first plotting data points. OLS regression similarly estimates how a dependent variable – say, RT Audience Score – is related to one or more independent variables – say, the five film measures with complete data.

Before we begin, however, let us set a baseline to see how well the estimation process worked: the correlation matrix derived from the 561 films with complete data. I characterize these correlations in the previous post, so I do not belabor them here.

Figure 1: Film Quality Measures Correlations Matrix (n=561)

Missing data estimation is an iterative process. I progressed from variables with the fewest missing cases (5 each for RT Audience Score and Average) to those with the most missing cases (59 for Tomatometer, RT Critic Average). Consecutive OLS regressions are summarized in Table 2; Intercooled Stata 9.2 (“Stata”)[2] was used for all analyses. Because Maltin Stars is an ordinal variable – it has seven discrete categories, rather than being continuous – I considered using ordinal logistic regression, but the OLS model proved a better fit; estimated Maltin Stars were rounded to the nearest “half-star.”

Table 2: Iterative Film Quality Measure OLS Regressions

VariableRT Aud ScoreRT Aud AveMaltin StarsRT Critic AveTomato- meter
Year of Release0.217-0.00005-0.007-0.0080.043
Length (mins.)0.0020.00060.004-0.002-0.033
IMDb Score25.0980.0370.4880.6771.576
IMDb Raters-0.000027.7e-8-8.8e-75.0e-8-4.4e-6
RT Critic Average    14.858
RT Critics0.001-0.00020.0020.002-0.017
RT Audience Score 0.022-0.0040.0250.434
RT Audience Average  0.298-0.268-19.504
Maltin Stars   0.5251.126
Number of cases633633615579579
Adjusted R-squared0.8090.9170.4610.7730.855

These are solid models, accounting for between 46.1 (Maltin Stars) and 91.7% (RT Audience Average) of a dependent variable’s variance. Moreover, the estimated values had high face validity – they made “sense.” And as Figure 2 shows, the 151 estimated cases barely changed relationships between these variables.

Figure 2: Film Quality Measures Correlations Matrix (n=638)

On average, the 45 correlations increased 0.045, with IMDb Score’s 9 correlations increasing a mean 0.013, and Runtime and Year of Release increasing by 0.086 and 0.134, respectively. The latter increases reflect the preponderance of missing data among both older/shorter and newer/longer films – SD increased, increasing all 18 correlations. The story is the same if you look at the absolute value of changes: an average shift of 0.050 in either direction, with Tomatometer’s 9 correlations changing by a mean of 0.020.

In other words, this data estimation process was very successful.


Armed with complete data, I used factor analysis[3] to calculate a single “film quality” score. The results were nearly identical to those from the previous post: using all 10 variables yields two factors accounting for 68.1% (PQ) and 26.0% (Public Awareness) of the total variance, while removing year of release, runtime and numbers of raters yields one factor (PQ) accounting for an astonishing 92.6% of the total variance: IMDb users, RT critics, RT fans and Maltin rate movies in remarkably similar ways. Table 3 presents the factor loadings (correlations with underlying dimension being assessed) and score coefficients used to generate a single PQ score.

Table 3: Factor Analysis of Film Quality Measures, Two Iterations

VariableAll 10 Variables 68.2%“Rating” Variables Only 92.6%
Year of Release-0.1720.013  
Length (mins.)0.2480.017  
IMDb Score0.9350.2270.9140.104
IMDb Raters0.4250.035  
RT Critic Average0.9210.3110.9320.339
RT Critics0.3910.059  
RT Audience Score0.9190.1940.9230.253
RT Audience Average0.9030.1710.9030.203
Maltin Stars0.6790.0380.6910.037

From these coefficients, Stata[4] calculated two PQ scores – PQAll and PQRating – for each film. Think of these values as SD above or below 9 to 5.  They are correlated a whopping 0.992, though they do have subtle differences, as Table 4 reveals:

Table 4: 30 Highest Rated Films I Have Seen Multiple Times, Compared by PQ Score

30. The Apartment30. The Maltese Falcon (1941)
29. Memento29. The Cabinet of Dr. Caligari
28. No Country For Old Men28. Witness for the Prosecution
27. Some Like It Hot27. L.A. Confidential
26. Sherlock Jr.26. The Apartment
25. Vertigo25. Vertigo
24. The Wizard of Oz24. Annie Hall
23. The Third Man23. Back to the Future
22. L.A. Confidential22. The General
21. The Avengers21. Kind Hearts and Coronets
20. Double Indemnity20. The Wizard of Oz
19. On the Waterfront19. Metropolis
18. M (1931)18. Some Like It Hot
17. Metropolis17. North By Northwest
16. North By Northwest16. Star Wars Episode IV: A New Hope
15. Chinatown15. Double Indemnity
14. Back to the Future14. Chinatown
13. Sunset Boulevard13. The Third Man
12. Rear Window12. On the Waterfront
11. Citizen Kane11. Rear Window
10. It’s A Wonderful Life10. It’s a Wonderful Life
9. Psycho9. M (1931)
8. Indiana Jones and the Raiders of the Lost Ark8. Psycho
7. Star Wars Episode IV: A New Hope7. Citizen Kane
6. Casablanca6. Indiana Jones and the Raiders of the Lost Ark
5. Inception5. Sunset Boulevard
4. 12 Angry Men4. Sherlock Jr.
3. Avengers: Endgame3. Pulp Fiction
2. Pulp Fiction2. Casablanca
1. The Dark Knight1. 12 Angry Men

Just 24 films appear on both lists (including L.A. Confidential, my favorite movie). Six newer (median 2009), longer (146.5 minutes), more-oft-rated films (1,214,741 IMDb raters, 352 RT critics) films in the PQAll Top 30 are replaced by six older (1945), shorter (96.5 minutes), slightly-less-oft-rated films (131,469; 58.5). Only six films – It’s A Wonderful Life, Psycho, Indiana Jones and the Raiders of the Lost Ark, Casablanca, 12 Angry Men and Pulp Fiction – rank in the top 10 on both PQ scores.

Still, if I were to choose a set of recent films likeliest still to be highly regarded a few decades from now, Memento, No Country for Old Men, The Avengers, Inception, Avengers: Endgame and The Dark Knight are an excellent starting point. Three of them, along with The Prestige and Batman Begins, were directed by Christopher Nolan – perhaps the best director of our time.

While this is a very impressive list of films – I was even more blown away by 12 Angry Men the second time – it is based ONLY on films I have seen multiple times. It excludes highly-regarded films (per the 20 top-ranked films on IMDb) I have only seen once: The Shawshank Redemption, first two Godfather films, Fight Club, Forrest Gump, Star Wars Episode V: The Empire Strikes Back, The Matrix, Goodfellas, One Flew Over the Cuckoo’s Nest and Se7en. And then there are movies I have not seen at all: Schindler’s List; the Lord of the Rings trilogy; The Good, the Bad and the Ugly; and Seven Samurai.

In fact, I have only seen four of the top 20 IMDb-rated movies multiple times: The Dark Knight, 12 Angry Men, Pulp Fiction and Inception. This reflects my personal taste in movies: older, noir-tinged, mysteries and comedies rather than more contemporary fantasy, war-based or western films; whereas the median year of release of the top 250 films by PQScore is 1971.5, respectively, that of the 250 films by IMDb score is 1994 – with 51 being released in 2011 or later, a strong indication of recency bias in the IMDb score data.

Another way to consider my particular taste in movies is to examine the distribution of year of release (Figure 3). There are two distinct peaks –1946-50 (n=73, 11.4%) and 1982-98 (n=256, 40.1%). The former period roughly corresponds to the pinnacle of classic film noir, while the latter is my primary movie-attending period – ages 15 to 32. Indeed, 119 (18.7%) of these films are classic-era films noir, released between 1940 (Stranger on the Third Floor) and 1958 (Touch of Evil) with average PQAll and PQScore of 0.03 and 0.08, respectively (relative to 0, overall). This excludes 18 films directed by Alfred Hitchcock, 10 of which are widely considered film noir, with average PQAll and PQScore of 0.92and 0.94, respectively. Finally, there are 49 films released before 1960 – not film noir or Hitchcock, not Charlie Chan (n=23; both -0.66), starring The Marx Brothers (n=9; 0.38, 0.48) – with averages of 0.92 and 0.97, respectively. The other 395 films (61.9%) – excluding the 25 directed by Woody Allen[5] (n=25; 0.26, 0.30) – have averages of -0.15 and -0.18, respectively. These values are broadly similar to those from the previous post, excepting the addition of “Charlie Chan.”

Figure 3: Distribution of Year of Release is Bimodal

Even more instructive is to compare my favorite film by time period (>5 films) – or, at least, my best guess absent a formal assessment – to the film with the highest PQAll and PQScore from that period.

1920 to 1930

Number films = 9

Average PQAll = 1.12

Average PQScore = 1.20

Top film: Metropolis (1.61, 1.53)

Personal favorite: The Phantom of the Opera (0.76, 0.84), albeit barely

Comment: Phantom has the lowest PQ scores of the nine films released in the 1920s I have seen multiple times, revealing their overall quality.


Number films = 5

Average PQAll = 0.65

Average PQScore = 0.74

Top film: M (1.61,1.65)

Personal favorite: M

Comment: I have seen at least five other films released in 1931, making it a key year in my personal fandom

1932 to 1934

Number films = 5

Average PQAll = 0.72

Average PQScore = 0.83

Top film: The Thin Man (1.32,1 .43)

Personal favorite: The Thin Man


Number films = 5

Average PQAll = 0.27

Average PQScore = 0.34

Top film: A Night at the Opera (1.23, 1.30)

Personal favorite: A Night at the Opera

1936 to 1939

Number films = 29

Average PQAll = -0.18

Average PQScore = -0.17

Top film: The Wizard of Oz (1.57, 1.50)

Personal favorites: After the Thin Man (0.83, 0.94), Charlie Chan at Treasure Island (-0.18, -0.15), Charlie Chan in Reno (-0.62, -0.64)

Comment: For me, the 1930s combine genuinely great films – eight from the Marx Brothers – with 16 Charlie Chan films.


Number films = 10

Average PQAll = 0.19

Average PQScore = 0.24

Top film: Rebecca (1.43, 1.41)

Personal favorite: Foreign Correspondent (0.66, 0.72), though Rebecca and His Girl Friday (1.28, 1.35) are wicked close

Comment: Hitchcock made his first Hollywood films in 1940, and they were knockouts…though the best was yet to come.


Number films = 11

Average PQAll = 0.19

Average PQScore = 0.23

Top film: Citizen Kane (1.75, 1.66)

Personal favorite: The Maltese Falcon (1.38, 1.44)

Comment: Film noir bursts on the scene with a bang.


Number films = 9

Average PQAll = 0.04

Average PQScore = 0.07

Top film: Casablanca (1.88, 1.77)

Personal favorite: All Through the Night (0.15, 0.26)

Comment: …as does Humphrey Bogart.

1943 to 1944

Number films = 10

Average PQAll = 0.64

Average PQScore = 0.70

Top film: Double Indemnity (1.61, 1.59)

Personal favorite: Laura (1.25, 1.31)

Comment: 1944 is the best year for film quality (0.74, 0.79) since 1931


Number films = 7

Average PQAll = -0.25

Average PQScore = -0.20

Top film: Scarlet Street (0.82, 0.91)

Personal favorite: Spellbound (0.40, 0.43)

Comment: Followed by a mediocre 1945.


Number films = 17

Average PQAll = 0.01

Average PQScore = 0.05

Top film = It’s A Wonderful Life (1.77, 1.65)

Personal favorite: Deadline at Dawn (-0.56, -0.50)

Comment: This is the first large disconnect between the “best” film released in a year and my favorite film from that year.


Number films = 19

Average PQAll = -0.18

Average PQScore = -0.22

Top film: Out of the Past (1.77, 1.65)

Personal favorite: Out of the Past


Number films = 13

Average PQAll = 0.30

Average PQScore = 0.38

Top film: Rope (0.91, 0.94)

Personal favorites: Call Northside 777 (0.09, 0.09) and The Naked City (0.39, 0.43)


Number films = 11

Average PQAll = 0.26

Average PQScore = 0.30

Top film = The Third Man (1.58, 1.62)

Personal favorite: Impact (-0.60, -0.65)


Number films = 13

Average PQAll = 0.08

Average PQScore = 0.12

Top film: Sunset Boulevard (1.70, 1.69)

Personal favorite: Where the Sidewalk Ends (0.59, 0.71)

Comment: Classic film noir and Hitchcock yielded some of the best films ever made.


Number films = 8

Average PQAll = 0.20

Average PQScore = 0.28

Top film: Strangers on a Train (1.27, 1.31)

Personal favorite: The Enforcer (-0.19, -0.15)


Number films = 6

Average PQAll = -0.05

Average PQScore = -0.01

Top film: The Narrow Margin (0.72, 0.85)

Personal favorite: Kansas City Confidential (-0.05, -0.02)

Comment: This is the first time the best film I have seen multiple times released that year is far from the best film released that year. That honor goes to Singin’ in the Rain, a movie I have yet to see in its entirety.


Number films = 7

Average PQAll = -0.02

Average PQScore = 0.02

Top film: The Big Heat (1.04, 1.12)

Personal favorite: 99 River Street (0.35, 0.47)


Number films = 7

Average PQAll = 0.40

Average PQScore = 0.38

Top film: Rear Window (1.72, 1.64)

Personal favorite: Dial M For Murder (0.88, 0.84), edging out Rear Window

Comment: I have yet to see Seven Samurai…and Hitchcock has hit his absolute peak.


Number films = 7

Average PQAll = 0.77

Average PQScore = 0.85

Top films: Diabolique (1.25, 1.27) and Rififi (1.24, 1.27)

Personal favorite: Muerte de un ciclista (Death of a Cyclist) (0.70, 0.82)

Comment: For the first time since 1920-31, foreign films dominate – with Germany replaced by France and Spain. It also marks the shift the overseas shift of film noir, presaging La Nouvelle Vague.

1956 to 1959

Number films = 15

Average PQAll = 0.94

Average PQScore = 0.94

Top film: 12 Angry Men (1.95, 1.80)

Personal favorite: 12 Angry Men


Number films = 26

Average PQAll = 0.31

Average PQScore = 0.33

Top film: Psycho (1.79, 1.66)

Personal favorite: The Apartment (1.49, 1.44), though Psycho is close.

Comment: I have not seen multiple times any films released in 1961, 1969 or 1970. And most of the 1960s is a cinematic wasteland for me…though I may have seen Dr. Strangelove, or How I Stopped Worrying and Learned to Love the Bomb more than once.

1970 to 1972

Number films = 6

Average PQAll = 0.36

Average PQScore = 0.39

Top film = Sleuth (1.17, 1.19)

Personal favorites: What’s Up, Doc? (0.70, 0.76) and Willie Wonka and the Chocolate Factory (0.79, 0.79)

Comment: The Godfather – which I have only seen once – is the best film released in 1972. Some would argue…ever.


Number films = 8

Average PQAll = 0.44

Average PQScore = 0.46

Top films:  Paper Moon (1.32, 1.36) and The Sting (1.37, 1.31)

Personal favorite: Charley Varrick (0.45, 0,49) edges The Sting and American Graffiti (0.83, 0.90).

1974 to 1975

Number films = 11

Average PQAll = 0.47

Average PQScore = 0.51

Top film:  Chinatown (1.66, 1.61)

Personal favorite: Murder on the Orient Express (0.41, 0.43)

Comment: Chinatown is a masterpiece; The Godfather: Part 2 (only seen once) is considered better.

1976 to 1977

Number films = 10

Average PQAll = 0.19

Average PQScore = 0.20

Top film: Star Wars Episode IV: A New Hope (1.87, 1.57)

Personal favorites: Murder by Death (-0.06, -0.05) and The Seven-Per-Cent Solution (-0.46, -0.39)


Number films = 11

Average PQAll = -0.64

Average PQScore = -0.61

Top film: Superman (0.81, 0.85)

Personal favorites: Death on the Nile (0.04, 0.03) and Thank God It’s Friday (-2.26, -2.24)

Comment: I genuinely do not understand why the charming Friday does get more love. And as solid as Superman is, The Deer Hunter (which I have not seen) is probably the best film released in 1978.


Number films = 12

Average PQAll = 0.06

Average PQScore = 0.11

Top films: Being There (1.20, 1.20) and Manhattan (1.19, 1.21)

Personal favorites: Manhattan and The In-Laws (0.50, 0.61)


Number films = 12

Average PQAll = -0.22

Average PQScore = -0.18

Top film: Airplane! (1.02, 1.06)

Personal favorite: Times Square (-0.60, -0.55)

Comment: I have only seen Empire Strikes Back once.


Number films = 11

Average PQAll = 0.23

Average PQScore = 0.29

Top film: Indiana Jones and the Raiders of the Lost Ark

Personal favorite: Indiana Jones and the Raiders of the Lost Ark


Number films = 17

Average PQAll = 0.06

Average PQScore = 0.09

Top film: Blade Runner (1.30, 1.14)

Personal favorites: Fast Times at Ridgemont High (0.13, 0.15) and Hammett (-0.60, -0.52)

Comment: Wow, Harrison Ford dominated movies from 1977 and 1983.


Number films = 9

Average PQAll = -0.18

Average PQScore = -0.12

Top film: A Christmas Story (1.14, 1.15)

Personal favorite: Valley Girl (-0.22, -0.06)


Number films = 19

Average PQAll = -0.34

Average PQScore = -0.30

Top film: This is Spinal Tap (1.22, 1.26)

Personal favorite: The Cotton Club (-0.67, -0.66)

Comment: We are now squarely in the age of mediocre films I saw in the theater as a teenager/young adult then chose to watch again. And maybe I need to see Amadeus again.


Number films = 14

Average PQAll = 0.22

Average PQScore = 0.24

Top film: Back to the Future (1.70, 1.48)

Personal favorite: Back to the Future, though The Sure Thing is close (0.09, 0.18)


Number films = 17

Average PQAll = -0.16

Average PQScore = -0.14

Top films: Hannah and Her Sisters (1.11, 1.12) and Blue Velvet (1.07, 1.08)

Personal favorite: While Legal Eagles (-1.73, -1.83) is a top guilty pleasure, I may like Hannah more.

Comment: Yes, I have never seen Aliens.


Number films = 17

Average PQAll = -0.33

Average PQScore = -0.32

Top film: Wings of Desire (1.39, 1.41)

Personal favorite: House of Games (0.57, 0.67)


Number films = 13

Average PQAll = -0.43

Average PQScore = -0.41

Top film: Die Hard (1.48, 1.33)

Personal favorite: Who Framed Roger Rabbit (0.97, 0.97)


Number films = 13

Average PQAll = -0.49

Average PQScore = -0.49

Top film: Crimes and Misdemeanors (0.98, 1.00)

Personal favorite: Forced to choose from a lot of meh, I pick The Big Picture (-0.84, -0.77) for its charming cast.


Number films = 13

Average PQAll = -0.32

Average PQScore = -0.29

Top film: Metropolitan (0.58, 0.66)

Personal favorite: Metropolitan

Comment: The one film released in 1990 in the IMDb Top 250 is Goodfellas, which I look forward to rewatching. [Ed. note: I neglected to add Awakenings, which I have seen twice, and which tops Metropolitan among films I have seen multiple times – though the latter is still my personal favorite among these group.]


Number films = 20

Average PQAll = -0.70

Average PQScore = -0.73

Top film: JFK (0.99, 0.85)

Personal favorite: If I have to pick one from this meh collection – Dead Again (0.15, 0.23)

Comment: I should watch The Silence of the Lambs again.


Number films = 16

Average PQAll = -0.63

Average PQScore = -0.64

Top film: The Player (0.94, 0.98)

Personal favorite: The Public Eye (-0.99, -1.01)

Comment: I should watch Reservoir Dogs again – as we hit rock bottom in the early 1990s.


Number films = 13

Average PQAll = -0.85

Average PQScore = -0.85

Top film: The Fugitive (1.01, 0.95)

Personal favorite: Manhattan Murder Mystery (0.34, 0.40)

Comment: I spoke too soon…yeesh. Perhaps once I finally see Schindler’s List. [Ed. note: A few days after posting this, I watched Dazed and Confused for the second time, and I think it replaces both The Fugitive and Manhattan Murder Mystery – making it both the “best” film and my favorite film release in 1993 I have seen multiple times. I have also seen Jurassic Park only once.]


Number films = 18

Average PQAll = -0.32

Average PQScore = -0.35

Top film: Pulp Fiction (2.16, 1.72)

Personal favorite: The Shadow (-1.63, -1.74) – easily the widest gap between “best” and “favorite”

Comment: Do not be fooled by these data. Hollywood has now entered an absolute golden age, rivaling its best years. Remember, I have only seen Shawshank Redemption, Forrest Gump and the phenomenal Léon: The Professional once. And I have never seen The Lion King. These four movies and Pulp Fiction are in the top 40 films by IMDb score.


Number films = 12

Average PQAll = -0.28

Average PQScore = -0.34

Top film: The Usual Suspects (2.16, 1.72)

Personal favorite: The Usual Suspects

Comment: Eight of the IMDb Top 250 were released in 1995.


Number films = 15

Average PQAll = -0.36

Average PQScore = -0.36

Top film: Fargo (1.42, 1.31)

Personal favorite: Big Night (0.69, 0.78), probably.


Number films = 9

Average PQAll = 0.24

Average PQScore = 0.20

Top film: L.A. Confidential (1.59, 1.44)

Personal favorite: L.A. Confidential

Comment: I am still upset L.A. Confidential did not win the Academy Award for Best Picture…and Titanic is not even in the IMDb Top 250; L.A. Confidential only ranks 5th.


Number films = 16

Average PQAll = -0.24

Average PQScore = -0.26

Top film: Rushmore (1.08, 1.07)

Personal favorites: Dark City (0.45, 0.37) and Pleasantville (0.54, 0.46)

Comment: I have zero interest in Saving Private Ryan – in fact, Steven Spielberg does almost nothing for me. There, I said it. And I just realized I left The Big Lebowski off my list – though it would not rank much higher than Rushmore.


Number films = 14

Average PQAll = -0.40

Average PQScore = -0.51

Top film: Toy Story 2 (1.34, 1.19)

Personal favorites: Cradle Will Rock (-0.24, -0.29) and Mystery Men (-0.91, -0.96)

Comment: I may watch Fight Club and The Matrix again at some point.

2000 to 2001

Number films = 12

Average PQAll = 0.35

Average PQScore = 0.21

Top film: Memento (1.52, 1.19)

Personal favorite: Mulholland Drive (1.03, 0.79)


Number films = 20

Average PQAll = 0.19

Average PQScore = -0.14

Top film: The Dark Knight (2.19, 1.39)

Personal favorite: The Dark Knight

Comment: For the first time, the difference in scores reflects the bias in PQAll toward longer, more recent films with many raters.


Number films = 12

Average PQAll = 1.09

Average PQScore = 0.72

Top film: Avengers: Endgame (2.03, 1.30) and Inception (1.95, 1.17)

Personal favorite: Predestination (0.28, 0.10)

Comment: We come full circle: as in the 1920s, the handful of films released in the 2010s I have seen multiple times are generally well-regarded. And my favorite (sorry, Endgame, Hugo, Doctor Strange) is again least well-regarded, but still better than average.


Figure 4 reinforces the year-by-year analysis. It shows early and late peaks: one for the 1920s because I carefully chose the best films to watch and rewatch, and one for the 2010s, because I stopped seeing movies in theatres first, choosing the best films to watch (OK, and all 23 Marvel Cinematic Universe films) then watch again. Moreover, average PQScore is higher than average PQAll until about 1940 – when it basically draws even until the early 1990s – because the latter elevates longer, recent, heavily-reviewed films. Around 2000, average PQAll pulls much further ahead for the same reason.

After a sharp decline in both values through the 1930s (excepting a slight uptick in 1931) – reflecting the Fox Charlie Chan films I love more than most – the emergence of film noir around 1940 pulls values up again. They generally stay above 0 through the early 1950s, peaking sharply in 1944. A preponderance of exceptional Hitchcock and foreign films sends scores skyrocketing in the late 1950s, with a lower peak from the 1960s into the mid-1970s: once again a lack of films elevates scores.

Figure 4: Average PQAll and PQScores by Year of Release/Midpoint of Range of Year of Release, 1920-2019

Almost by definition, I first saw any film released through the mid-1970s, not in a movie theater, but on television, through one of Yale’s films societies or through a rental/streaming service. This will inevitably bias toward “better” films. That changed in the late 1970s, when I began regularly seeing films for the first time in a movie theater – resulting in a much wider range of quality. The steep decline in scores in 1978 shows that, as does the zig-zagging around 0 through the mid-1980s. I cannot really explain why scores plummet in the late 1980s and early 1990s, though there is evidence overall film quality was much lower in those year.

The sharp spike up in 1994 masks an even more dramatic increase in overall film quality. I have long thought the peace and prosperity of the last six years of the Clinton Administration yielded a new golden age in American cinema; perhaps studio executives felt freer to experiment with original screenplays. I also though that changed after the 9-11, but…maybe not. Still, I suspect the preponderance of post-2000 films in the IMDb Top 250 reflects recency bias more than actual quality.

Basically, while I consider myself a cinephile, like everyone who watches movies my tastes range from genuine works of art to the guiltiest of guilty pleasures. And that is how it is supposed to be – we like what we like, not what we are “supposed” to like. Still, whether I am simply more honest in admitting how much I genuinely like the Fox Charlie Chan films, and films like Deadline at Dawn, Impact, The Seven-Per-Cent Solution, Thank God It’s Friday, Times Square, Hammett, The Cotton Club, Legal Eagles, The Big Picture, The Public Eye, The Shadow and Mystery Men remains to be seen.

In the meantime, I just shared Cat People (1942) with my wife Nell – who quite enjoyed it – so I need to update my Excel workbook. Again.

Until next time…

[1] New York, NY: New American Library

[2] StataCorp. 2005. Stata Statistical Software: Release 9. College Station, TX: StataCorp LP.

[3] Specifically: factor analysis, principal factors, varimax rotation, forcing one or two factors, depending on input variables.

[4] Using “Predict” command in Stata

[5] Despite my ambivalence about Allen as a human being, I still love many of his films.